There are many algorithms and no silver bullet - the performance of the algorithm also depends on the provided dataset. RBM is known to traditionally work well with Netflix data see Netflix blog: http://techblog.netflix.com/2012/04/netflix-recommendations-beyond-5-stars.html
When Netflix talks about SVD they do not mean Singular Value Decomposition, but SVD++ which is also implemented in GraphChi.
You should note that the performance of the algorithm depends on many tunable parameters, and you should optimize the number of iterations, feature vector width (D), regularization and special parameters for each algorithm for example step sizes for SGD. I believe the results you sent are not optimized yet but present a baseline.